ICCV Daily 2021 - Friday

Image classification has many practical applications but applying it at scale is difficult because of a lack of large, labeled datasets . Creating new labeled datasets takes a lot of time and effort, and labels can be noisy. It is easier to scale to much larger models and better classifiers if you can use unlabeled data . “ You need to find a way to work when you don’t have a lot of labeled data, ” Mahmoud tells us. “ It’s really easy to get unlabeled data, but in many applications, it’s hard to get the labeled data, so we have to find a way to work with this. ” In this work, Mahmoud and his colleagues propose an image classification algorithm called PAWS which is trained on a large set of unlabeled images and a very small set of labeled images . The challenging part has been beating existing benchmarks in semi- supervised image classification because it is such a dense field. A few months before this work, another paper came out of Google Brain with very strong benchmarks. The field is constantly evolving and moves very fast. Mahmoud solved this by taking inspiration from all the progress in self- supervised learning and integrating some of those ideas into a semi- supervised framework . Semi-supervised methods before PAWS have had a 8 DAILY ICCV Friday Oral Presentation Semi-Supervised Learning of Visual Features by Non-Parametrically Predicting View Assignments with Support Samples Mahmoud (Mido) Assran is a PhD student at McGill University and Mila, the Quebec Artificial Intelligence Institute. He is also a visiting researcher at Facebook AI Research in the Montreal lab. Mahmoud is first author on a paper about self-supervised image classification which has been accepted as an oral presentation. He tells us more ahead of his live Q&A session today.